https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118120
--- Comment #10 from kargls at comcast dot net ---
(In reply to anlauf from comment #9)
> (In reply to kargls from comment #8)
> > (In reply to anlauf from comment #7)
> > > The following patch works and might be a reasonable compromise:
> > >
> > > diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc
> > > index 82a2ae1f747..985a26281ad 100644
> > > --- a/gcc/fortran/trans-array.cc
> > > +++ b/gcc/fortran/trans-array.cc
> > > @@ -5345,7 +5345,10 @@ symbols_could_alias (gfc_symbol *lsym, gfc_symbol
> > > *rsym, bool lsym_pointer,
> > > bool lsym_target, bool rsym_pointer, bool
> > > rsym_target)
> > > {
> > > /* Aliasing isn't possible if the symbols have different base types.
> > > */
See below. You need to update the comment to reflect that the symbols
can have different base types.
> > > - if (gfc_compare_types (&lsym->ts, &rsym->ts) == 0)
> > > + if (!gfc_compare_types (&lsym->ts, &rsym->ts)
> > > + && !(((lsym->ts.type == BT_COMPLEX && rsym->ts.type == BT_REAL)
> > > + || (lsym->ts.type == BT_REAL && rsym->ts.type == BT_COMPLEX))
> > > + && lsym->ts.kind == rsym->ts.kind))
> > > return 0;
> > >
> > > /* Pointers can point to other pointers and target objects. */
> >
> > I think it is more complicated in that this involves pointers, which
> > have an aliasing issue (due to bounds remapping). In an ordinary assignment
> > one has 'var = expr', and in Slava example that is exactly what we have
> > where 'result' and 'temp' are pointers into parts of 'data%re'. But,
> > 'data%re' could have been any type compatible entity. The big hammer,
> > until the mythical someone shows up, is
> >
> > diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc
> > index 82a2ae1f747..abebf327621 100644
> > --- a/gcc/fortran/trans-array.cc
> > +++ b/gcc/fortran/trans-array.cc
> > @@ -5503,6 +5503,20 @@ gfc_conv_resolve_dependencies (gfc_loopinfo * loop,
> > gfc_ss * dest,
> > loop->temp_ss = NULL;
> > dest_expr = dest->info->expr;
> >
> > + /* Before falling into dependency checking of the scalarization,
> > + check to see if the left-hand variable is an array pointer. If
> > + it is, simply generate an array temporary for the assignment.
> > + TODO: walk the right-hand-side of 'var = expr' to actually check
> > + for a dependency. See 118120. */
> > + if (dest_expr->symtree && dest_expr->symtree->n.sym
> > + && dest_expr->symtree->n.sym->attr.pointer
> > + && dest_expr->symtree->n.sym->attr.dummy
> > + && dest_expr->symtree->n.sym->attr.dimension)
> > + {
> > + nDepend = 1;
> > + goto temporary;
> > + }
> > +
> > for (ss = rss; ss != gfc_ss_terminator; ss = ss->next)
> > {
> > ss_info = ss->info;
>
> I doubt that this is a good solution.
>
> It is completely irrelevant that "result" is a dummy. Just try it.
> And creating a temporary for *every lhs pointer* cannot be acceptable.
> There must be something more "lightweight".
I agree it's not the best solution. This is why I called it
a 'big hammer'. Checking for a dummy argument is simply
trying to reduce the biggest of the hammer. I suspect the
aliasing issue is more likely to occur through argument
association, but users can be creative. Feel free to
pursue your patch with an expanded comment.
